[PATCH] Fix REE for vector modes (PR rtl-optimization/64286)

[PATCH] Fix REE for vector modes (PR rtl-optimization/64286)

[Jakub Jelinek]

Hi!

As mentioned in the PR, giving up for all vector mode extensions
is unnecessary, but unlike scalar integer extensions, where the low part
of the extended value is the original value, for vectors this is not true,
thus the old value is lost.  Which means we can perform REE, but only if
all uses of the definition are the same (code+mode) extension.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2015-01-12  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/64286
	* ree.c (add_removable_extension): Don't add vector mode
	extensions if all uses of the source register aren't the same
	vector extensions.

	* gcc.target/i386/avx2-pr64286.c: New test.

--- gcc/ree.c.jj	2015-01-09 22:00:00.427660442 +0100
+++ gcc/ree.c	2015-01-12 12:15:56.097184674 +0100
@@ -1027,6 +1027,7 @@ add_removable_extension (const_rtx expr,
 	 different extension.  FIXME: this obviously can be improved.  */
       for (def = defs; def; def = def->next)
 	if ((idx = def_map[INSN_UID (DF_REF_INSN (def->ref))])
+	    && idx != -1U
 	    && (cand = &(*insn_list)[idx - 1])
 	    && cand->code != code)
 	  {
@@ -1038,6 +1039,57 @@ add_removable_extension (const_rtx expr,
 	      }
 	    return;
 	  }
+	/* For vector mode extensions, ensure that all uses of the
+	   XEXP (src, 0) register are the same extension (both code
+	   and to which mode), as unlike integral extensions lowpart
+	   subreg of the sign/zero extended register are not equal
+	   to the original register, so we have to change all uses or
+	   none.  */
+	else if (VECTOR_MODE_P (GET_MODE (XEXP (src, 0))))
+	  {
+	    if (idx == 0)
+	      {
+		struct df_link *ref_chain, *ref_link;
+
+		ref_chain = DF_REF_CHAIN (def->ref);
+		for (ref_link = ref_chain; ref_link; ref_link = ref_link->next)
+		  {
+		    if (ref_link->ref == NULL
+			|| DF_REF_INSN_INFO (ref_link->ref) == NULL)
+		      {
+			idx = -1U;
+			break;
+		      }
+		    rtx_insn *use_insn = DF_REF_INSN (ref_link->ref);
+		    const_rtx use_set;
+		    if (use_insn == insn || DEBUG_INSN_P (use_insn))
+		      continue;
+		    if (!(use_set = single_set (use_insn))
+			|| !REG_P (SET_DEST (use_set))
+			|| GET_MODE (SET_DEST (use_set)) != GET_MODE (dest)
+			|| GET_CODE (SET_SRC (use_set)) != code
+			|| !rtx_equal_p (XEXP (SET_SRC (use_set), 0),
+					 XEXP (src, 0)))
+		      {
+			idx = -1U;
+			break;
+		      }
+		  }
+		if (idx == -1U)
+		  def_map[INSN_UID (DF_REF_INSN (def->ref))] = idx;
+	      }
+	    if (idx == -1U)
+	      {
+		if (dump_file)
+		  {
+		    fprintf (dump_file, "Cannot eliminate extension:\n");
+		    print_rtl_single (dump_file, insn);
+		    fprintf (dump_file,
+			     " because some vector uses aren't extension\n");
+		  }
+		return;
+	      }
+	  }
 
       /* Then add the candidate to the list and insert the reaching definitions
          into the definition map.  */
--- gcc/testsuite/gcc.target/i386/avx2-pr64286.c.jj	2015-01-12 12:19:54.863031657 +0100
+++ gcc/testsuite/gcc.target/i386/avx2-pr64286.c	2015-01-12 12:19:36.000000000 +0100
@@ -0,0 +1,37 @@
+/* PR rtl-optimization/64286 */
+/* { dg-do run } */
+/* { dg-options "-O2 -mavx2" } */
+/* { dg-require-effective-target avx2 } */
+
+#include <string.h>
+#include <stdlib.h>
+#include <x86intrin.h>
+#include "avx2-check.h"
+
+__m128i v;
+__m256i w;
+
+__attribute__((noinline, noclone)) void
+foo (__m128i *p, __m128i *q)
+{
+  __m128i a = _mm_loadu_si128 (p);
+  __m128i b = _mm_xor_si128 (a, v);
+  w = _mm256_cvtepu8_epi16 (a);
+  *q = b;
+}
+
+static void
+avx2_test (void)
+{
+  v = _mm_set1_epi8 (0x40);
+  __m128i c = _mm_set_epi8 (16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1);
+  __m128i d;
+  foo (&c, &d);
+  __m128i e = _mm_set_epi8 (0x50, 0x4f, 0x4e, 0x4d, 0x4c, 0x4b, 0x4a, 0x49,
+			    0x48, 0x47, 0x46, 0x45, 0x44, 0x43, 0x42, 0x41);
+  __m256i f = _mm256_set_epi16 (16, 15, 14, 13, 12, 11, 10, 9,
+				8, 7, 6, 5, 4, 3, 2, 1);
+  if (memcmp (&w, &f, sizeof (w)) != 0
+      || memcmp (&d, &e, sizeof (d)) != 0)
+    abort ();
+}

	Jakub

Re: [PATCH] Fix REE for vector modes (PR rtl-optimization/64286)

[Jeff Law]

On 01/12/15 12:59, Jakub Jelinek wrote:
> Hi!
>
> As mentioned in the PR, giving up for all vector mode extensions
> is unnecessary, but unlike scalar integer extensions, where the low part
> of the extended value is the original value, for vectors this is not true,
> thus the old value is lost.  Which means we can perform REE, but only if
> all uses of the definition are the same (code+mode) extension.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2015-01-12  Jakub Jelinek  <jakub@redhat.com>
>
> 	PR rtl-optimization/64286
> 	* ree.c (add_removable_extension): Don't add vector mode
> 	extensions if all uses of the source register aren't the same
> 	vector extensions.
>
> 	* gcc.target/i386/avx2-pr64286.c: New test.
Does it make sense to remove your change for 59754 in 
combine_reaching_defs?  Shouldn't this patch handle that case as well?

jeff

[PATCH] Fix REE for vector modes (PR rtl-optimization/64286, take 2)

[Jakub Jelinek]

On Mon, Jan 12, 2015 at 02:29:53PM -0700, Jeff Law wrote:
> On 01/12/15 12:59, Jakub Jelinek wrote:
> >Hi!
> >
> >As mentioned in the PR, giving up for all vector mode extensions
> >is unnecessary, but unlike scalar integer extensions, where the low part
> >of the extended value is the original value, for vectors this is not true,
> >thus the old value is lost.  Which means we can perform REE, but only if
> >all uses of the definition are the same (code+mode) extension.
> >
> >Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
> >
> >2015-01-12  Jakub Jelinek  <jakub@redhat.com>
> >
> >	PR rtl-optimization/64286
> >	* ree.c (add_removable_extension): Don't add vector mode
> >	extensions if all uses of the source register aren't the same
> >	vector extensions.
> >
> >	* gcc.target/i386/avx2-pr64286.c: New test.
> Does it make sense to remove your change for 59754 in combine_reaching_defs?
> Shouldn't this patch handle that case as well?

You're right, this patch handles that too.

New patch, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?

2015-01-13  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/64286
	* ree.c (combine_reaching_defs): Move part of comment earlier,
	remove !SCALAR_INT_MODE_P check.
	(add_removable_extension): Don't add vector mode
	extensions if all uses of the source register aren't the same
	vector extensions.

	* gcc.target/i386/avx2-pr64286.c: New test.

--- gcc/ree.c.jj	2015-01-12 21:29:07.023060045 +0100
+++ gcc/ree.c	2015-01-13 09:43:32.158449885 +0100
@@ -783,6 +783,17 @@ combine_reaching_defs (ext_cand *cand, c
        != REGNO (get_extended_src_reg (SET_SRC (PATTERN (cand->insn)))));
   if (copy_needed)
     {
+      /* Considering transformation of
+	 (set (reg1) (expression))
+	 ...
+	 (set (reg2) (any_extend (reg1)))
+
+	 into
+
+	 (set (reg2) (any_extend (expression)))
+	 (set (reg1) (reg2))
+	 ...  */
+
       /* In theory we could handle more than one reaching def, it
 	 just makes the code to update the insn stream more complex.  */
       if (state->defs_list.length () != 1)
@@ -798,18 +809,6 @@ combine_reaching_defs (ext_cand *cand, c
       if (state->modified[INSN_UID (cand->insn)].kind != EXT_MODIFIED_NONE)
 	return false;
 
-      /* Transformation of
-	 (set (reg1) (expression))
-	 (set (reg2) (any_extend (reg1)))
-	 into
-	 (set (reg2) (any_extend (expression)))
-	 (set (reg1) (reg2))
-	 is only valid for scalar integral modes, as it relies on the low
-	 subreg of reg1 to have the value of (expression), which is not true
-	 e.g. for vector modes.  */
-      if (!SCALAR_INT_MODE_P (GET_MODE (SET_DEST (PATTERN (cand->insn)))))
-	return false;
-
       machine_mode dst_mode = GET_MODE (SET_DEST (PATTERN (cand->insn)));
       rtx src_reg = get_extended_src_reg (SET_SRC (PATTERN (cand->insn)));
 
@@ -1027,6 +1026,7 @@ add_removable_extension (const_rtx expr,
 	 different extension.  FIXME: this obviously can be improved.  */
       for (def = defs; def; def = def->next)
 	if ((idx = def_map[INSN_UID (DF_REF_INSN (def->ref))])
+	    && idx != -1U
 	    && (cand = &(*insn_list)[idx - 1])
 	    && cand->code != code)
 	  {
@@ -1038,6 +1038,57 @@ add_removable_extension (const_rtx expr,
 	      }
 	    return;
 	  }
+	/* For vector mode extensions, ensure that all uses of the
+	   XEXP (src, 0) register are the same extension (both code
+	   and to which mode), as unlike integral extensions lowpart
+	   subreg of the sign/zero extended register are not equal
+	   to the original register, so we have to change all uses or
+	   none.  */
+	else if (VECTOR_MODE_P (GET_MODE (XEXP (src, 0))))
+	  {
+	    if (idx == 0)
+	      {
+		struct df_link *ref_chain, *ref_link;
+
+		ref_chain = DF_REF_CHAIN (def->ref);
+		for (ref_link = ref_chain; ref_link; ref_link = ref_link->next)
+		  {
+		    if (ref_link->ref == NULL
+			|| DF_REF_INSN_INFO (ref_link->ref) == NULL)
+		      {
+			idx = -1U;
+			break;
+		      }
+		    rtx_insn *use_insn = DF_REF_INSN (ref_link->ref);
+		    const_rtx use_set;
+		    if (use_insn == insn || DEBUG_INSN_P (use_insn))
+		      continue;
+		    if (!(use_set = single_set (use_insn))
+			|| !REG_P (SET_DEST (use_set))
+			|| GET_MODE (SET_DEST (use_set)) != GET_MODE (dest)
+			|| GET_CODE (SET_SRC (use_set)) != code
+			|| !rtx_equal_p (XEXP (SET_SRC (use_set), 0),
+					 XEXP (src, 0)))
+		      {
+			idx = -1U;
+			break;
+		      }
+		  }
+		if (idx == -1U)
+		  def_map[INSN_UID (DF_REF_INSN (def->ref))] = idx;
+	      }
+	    if (idx == -1U)
+	      {
+		if (dump_file)
+		  {
+		    fprintf (dump_file, "Cannot eliminate extension:\n");
+		    print_rtl_single (dump_file, insn);
+		    fprintf (dump_file,
+			     " because some vector uses aren't extension\n");
+		  }
+		return;
+	      }
+	  }
 
       /* Then add the candidate to the list and insert the reaching definitions
          into the definition map.  */
--- gcc/testsuite/gcc.target/i386/avx2-pr64286.c.jj	2015-01-13 09:33:53.986612033 +0100
+++ gcc/testsuite/gcc.target/i386/avx2-pr64286.c	2015-01-13 09:33:53.986612033 +0100
@@ -0,0 +1,37 @@
+/* PR rtl-optimization/64286 */
+/* { dg-do run } */
+/* { dg-options "-O2 -mavx2" } */
+/* { dg-require-effective-target avx2 } */
+
+#include <string.h>
+#include <stdlib.h>
+#include <x86intrin.h>
+#include "avx2-check.h"
+
+__m128i v;
+__m256i w;
+
+__attribute__((noinline, noclone)) void
+foo (__m128i *p, __m128i *q)
+{
+  __m128i a = _mm_loadu_si128 (p);
+  __m128i b = _mm_xor_si128 (a, v);
+  w = _mm256_cvtepu8_epi16 (a);
+  *q = b;
+}
+
+static void
+avx2_test (void)
+{
+  v = _mm_set1_epi8 (0x40);
+  __m128i c = _mm_set_epi8 (16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1);
+  __m128i d;
+  foo (&c, &d);
+  __m128i e = _mm_set_epi8 (0x50, 0x4f, 0x4e, 0x4d, 0x4c, 0x4b, 0x4a, 0x49,
+			    0x48, 0x47, 0x46, 0x45, 0x44, 0x43, 0x42, 0x41);
+  __m256i f = _mm256_set_epi16 (16, 15, 14, 13, 12, 11, 10, 9,
+				8, 7, 6, 5, 4, 3, 2, 1);
+  if (memcmp (&w, &f, sizeof (w)) != 0
+      || memcmp (&d, &e, sizeof (d)) != 0)
+    abort ();
+}


	Jakub

Re: [PATCH] Fix REE for vector modes (PR rtl-optimization/64286, take 2)

[Jeff Law]

On 01/13/15 09:11, Jakub Jelinek wrote:
> On Mon, Jan 12, 2015 at 02:29:53PM -0700, Jeff Law wrote:
>> On 01/12/15 12:59, Jakub Jelinek wrote:
>>> Hi!
>>>
>>> As mentioned in the PR, giving up for all vector mode extensions
>>> is unnecessary, but unlike scalar integer extensions, where the low part
>>> of the extended value is the original value, for vectors this is not true,
>>> thus the old value is lost.  Which means we can perform REE, but only if
>>> all uses of the definition are the same (code+mode) extension.
>>>
>>> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>>>
>>> 2015-01-12  Jakub Jelinek  <jakub@redhat.com>
>>>
>>> 	PR rtl-optimization/64286
>>> 	* ree.c (add_removable_extension): Don't add vector mode
>>> 	extensions if all uses of the source register aren't the same
>>> 	vector extensions.
>>>
>>> 	* gcc.target/i386/avx2-pr64286.c: New test.
>> Does it make sense to remove your change for 59754 in combine_reaching_defs?
>> Shouldn't this patch handle that case as well?
>
> You're right, this patch handles that too.
>
> New patch, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk?
>
> 2015-01-13  Jakub Jelinek  <jakub@redhat.com>
>
> 	PR rtl-optimization/64286
> 	* ree.c (combine_reaching_defs): Move part of comment earlier,
> 	remove !SCALAR_INT_MODE_P check.
> 	(add_removable_extension): Don't add vector mode
> 	extensions if all uses of the source register aren't the same
> 	vector extensions.
>
> 	* gcc.target/i386/avx2-pr64286.c: New test.
OK.  Thanks for taking care of this.  I can't seem to find time for 
doing any real debugging or bugfixing.

jeff